詞語驗證是語音辨識重要的後處理程序，目的在過濾掉可能是錯誤的辨識結果，提高語音辨 識的可靠度。而在硬體不斷朝向輕、薄、短、小的演化趨勢下，如何在有限的運算資源與記憶體 空間下，提供效果良好的詞語驗證作法，是當今具備語音辨識功能的電子產品所關注的重點。本 文提出一種簡易又有效的詞語驗證作法，應用語音辨識過程中所產生的辨識分數，以最大似然性 (Maximal Likelihood) 與最佳路徑似然性差異程度為驗證基準，並在不需額外付出大量的儲存空間與運算下，來改良資源受限平台上的詞語驗證效果。

Utterance verification is an important post-processing module for filtering out the probably wrong
hypotheses of recognition results, therefore ensuring the reliability of speech recognition system. In the
revolutionary trends with light, thin, short, and small hardware features, an effective utterance verification
module for resource-limited platform is necessary for modern electronics with speech recognition function. In
this paper, a simplified and effective method for utterance verification is proposed. The difference between t he
maximal likelihood and the optimal likelihood, generated in the speech recognition process, acts as the
verification basis. Without much extra memory and computations, the verification performance is improved by
using the proposed basis.